Patent Valuation under Fragile Institutional Enforcement: A Continuous-Time Markov Approach

Environmental, Social, and Governance (ESG) considerations have become integral to corporate strategy, investor decision-making, and regulatory oversight. ESG violations—such as environmental harm, governance failures, and social misconduct—pose substantial reputational, financial, and legal risks. This study develops a machine learning-based framework for the early detection of ESG policy violations using the World Benchmarking Alliance’s Nature Benchmark dataset (2022–2024), covering 816 firms across more than 20 industries. To address the pronounced class imbalance inherent in ESG violation data, the Synthetic Minority Over-sampling Technique (SMOTE) is applied. Three classification models—Logistic Regression, Decision Tree, and Random Forest—are evaluated. The Random Forest model demonstrates the most robust performance, achieving a superior balance between accuracy and recall. Model interpretability is ensured through feature importance measures and SHAP values, which identify key ESG dimensions and industry-specific drivers associated with violations. Overall, the findings highlight the effectiveness of combining ensemble learning, resampling techniques, and explainable machine learning to support scalable and proactive ESG risk assessment.